A spatio-temporal graph neural network with masked self-supervision for precise anomaly severity measurement of vibrating screens in mineral processing

Abstract

Accurately quantifying the anomaly severity of vibrating screens is crucial for ensuring the stable operation of mining production systems such as coal preparation plants. However, this task faces three major challenges: (1) Traditional methods struggle to learn robust health state baselines from normal operating data; (2) Vibration signals exhibit multi-scale temporal patterns and dynamic coupling relationships between sensors, posing difficulties for spatiotemporal feature modelling; (3) There is a lack of effective anomaly severity quantification mechanisms. To address these issues, this paper proposes a Spatio-Temporal Graph Neural Network with Masked Self-Supervision (STGNN-MSS). The method tackles the aforementioned challenges through three modules: First, the self-supervised pre-training module employs a Transformer-based masked autoencoder to learn deep temporal representations from normal data through a regression-classification dual-task framework; Second, the spatio-temporal feature learning module combines multi-scale hypergraph networks and dynamic graph convolutional networks to capture high-order dependencies at different temporal scales and dynamic spatial coupling between sensors, respectively; Finally, the anomaly severity detection module adopts a prediction-reconstruction dual-task framework and introduces a KL divergence-based distribution deviation metric to achieve precise quantification of anomaly severity. Experimental results demonstrate that STGNN-MSS achieves an F1 score of 0.9798 and an AUC value of 0.94 on the vibrating screen dataset, representing improvements of 11.2 % and 5.6 % over the best baseline methods, respectively, validating the effectiveness of the proposed method.